KR20220160407A - Device and method for predicting biomedical association - Google Patents

Abstract

According to one embodiment of the present invention, a method for predicting biomedical association performed by a computing device may include: a step of extracting real world data including a diagnosis code connected with patient information and a prescription code from a real world database; a step of extracting biomedical data including genetic information, a disease identifier, a chemical identifier, and a maternal structure identifier from a biomedical database; a step of constructing a new biological network by coupling the real world data and the biomedical data; a step of training a machine learning model by using random work data generated based on random work from the constructed new biological network; and a step of predicting a link between object nodes based on an embedding vector of an object node of a new biological network based on the trained machine learning model. According to one embodiment of the present invention, the computing device may construct a database that can be expanded according to the addition of a biomedical database.

Description

Method and apparatus for predicting biomedical relevance {DEVICE AND METHOD FOR PREDICTING BIOMEDICAL ASSOCIATION}

Hereinafter, a description of a method and apparatus for predicting biomedical relevance is provided.

With the development of electronic devices and the generalization of systematic electronic record collection, patient clinical data is being accumulated in the form of an Electronic Medical Record (EMR). EMR is not clinical data in a variable-controlled test environment, but data in the real world in which various variables are complexly included, and corresponds to Real World Data (RWD). Clinical evidence about the use and potential benefits and harms of medicines that can be obtained by analyzing RWD is called Real World Evidence (RWE). In recent clinical studies, attempts to predict or explain patients' biological phenomena or actual clinical events based on RWE are being studied from various angles.

Research is being conducted to explain the complex life activities of diseases and to predict the relationship between diseases and drugs. In various fields such as molecular biology and genomics, open biomedical databases for each related substance have been established and are used for research to understand the complex phenomena of life activities. Biomedical database-based relational research is concerned with the interaction of related elements such as genes, proteins, organs, drugs, and diseases, and the large context they form. ) can be a new approach to understanding

The background art described above is possessed or acquired by the inventor in the process of deriving the disclosure of the present application, and cannot necessarily be said to be known art disclosed to the general public prior to the present application.

A computing device according to an embodiment may combine biological and chemical knowledge accumulated in a biomedical database with clinical knowledge of an electronic medical record.

A computing device according to an embodiment may search for biomedical relevance through graph embedding technology in a database constructed in combination with a graph structure.

A computing device according to an embodiment may build a database by adding a patient and a table of an electronic medical record.

A computing device according to an embodiment may build a database according to the addition of a biomedical database.

However, the technical challenges are not limited to the above-described technical challenges, and other technical challenges may exist.

A biomedical correlation prediction method performed by a computing device according to an embodiment includes extracting real world data including a diagnosis code and a prescription code connected to patient information from a real world database; extracting biomedical data including genetic information, disease identifiers, chemical substance identifiers, and parent structure identifiers from a biomedical database; constructing a new biological network by combining the real world data and the biomedical data; training a machine learning model from the constructed new biological network; and predicting links between entity nodes based on an embedding vector of entity nodes of the new biological network based on the trained machine learning model.

The extracting of the real world data may include connecting a patient node indicating patient information to a prescription node indicating a prescription code extracted corresponding to the patient information and a diagnosis node indicating a diagnosis code through an edge, resulting in a graph structure It may include generating the real world data of.

The extracting of the biomedical data may include connecting a gene node indicating genetic information to a disease node indicating a disease identifier corresponding to the genetic information through an edge; connecting the disease node to a chemical node indicating a chemical identifier corresponding to the disease identifier through an edge; connecting the chemical node to a parent structure node indicating a parent structure identifier corresponding to the chemical identity through an edge; and generating biomedical data having a graph structure including the gene node, the disease node, the chemical node, and the parent structure node connected through an edge.

In the biomedical data, the disease node may be connected to the parent structure node via the chemical node through an edge, and the chemical node may be connected to the gene node via the disease node through an edge.

The step of constructing the new biological network may include connecting a diagnosis node of the real world data and a disease node of the biomedical data, and connecting a prescription node of the real world data and a chemical node of the biomedical data to form the new biological network. It may include building steps.

The constructing of the new biological network may include mapping the diagnosis node and the disease node corresponding to the same unique concept identifier based on a common data model for biomedical terminology; and connecting the prescription node and the chemical node corresponding to the same unique concept identifier based on the common data model.

The constructing of the new biological network may include extracting a first string corresponding to a diagnosis code indicated by the diagnosis node; Retrieving a first concept unique identifier having the extracted first string from a common data model; and mapping the diagnosis node to the disease node indicating a disease identifier corresponding to the retrieved first concept unique identifier.

The constructing of the new biological network may include extracting a second string corresponding to a prescription code indicated by the prescription node; Retrieving a second concept unique identifier corresponding to the extracted second string from a common data model; and mapping the prescription node to the chemical substance node indicating a chemical substance identifier corresponding to the retrieved second concept unique identifier.

The step of training the machine learning model may include selecting an initial entity node from the constructed new biological network; selecting entity nodes according to individual random walks by sequentially performing random walks as long as a predetermined random walk length from the selected initial entity nodes; and generating random walk data indicating a random walk set including the initial entity node and entity nodes selected based on the random walk.

The step of training the machine learning model may include calculating temporary embedding data by applying the machine learning model to input data based on the random walk data; and updating parameters of a machine learning model to maximize a connection probability between individual nodes in the random walk set based on the calculated temporary embedding data.

The step of predicting a link between entity nodes may include extracting a first embedding vector corresponding to a first node among entity nodes of the new biological network from the machine learning model; extracting a second embedding vector corresponding to a second node among entity nodes of the new biological network from the machine learning model; and determining whether an edge exists between the first node and the second node based on the extracted first embedding vector and the extracted second embedding vector.

In the step of determining whether an edge exists, an edge exists between the first node and the second node in response to a case where a vector distance difference between the extracted first embedding vector and the second embedding vector is less than a threshold distance value. It may include the step of deciding to do it.

A computing device according to an embodiment includes a memory storing a machine learning model for biomedical correlation prediction; And extracting real world data including diagnosis codes and prescription codes linked to patient information from real world databases, and extracting biomedical data including genetic information, disease identifiers, chemical substance identifiers, and parent structure identifiers from biomedical databases, A new biological network is constructed by combining the real world data and the biomedical data, a machine learning model is trained using random walk data generated based on a random walk from the new biological network, and the trained machine and a processor for predicting links between entity nodes based on an embedding vector of entity nodes of the new biological network based on a learning model.

A computing device according to an embodiment may effectively combine biological and chemical knowledge accumulated in a biomedical database with clinical knowledge of an electronic medical record.

The computing device according to an embodiment can effectively search for correlations between biological entities, such as drug indications, drug side effects, and drug treatment effects, through graph embedding technology in a database constructed in a graph structure.

The computing device according to an embodiment may build an expandable database according to the addition of patients and tables of the electronic medical record.

A computing device according to an embodiment may build a database that is expandable according to the addition of a biomedical database.

1 shows a block diagram of a biomedical correlation prediction device according to an embodiment.
2 is a flowchart illustrating a biomedical correlation prediction method according to an embodiment.
3 may represent real world data according to an embodiment.
4 may represent biomedical data according to an embodiment.
5 to 7 illustrate mapping of real world data and biomedical data according to an embodiment.
8 illustrates a mapping result between real world data and biomedical data according to an embodiment.
9 and 10 illustrate a new merged network created according to one embodiment.
11 illustrates training of a machine learning model and biomedical correlation using the trained machine learning model according to an embodiment.
12 illustratively illustrates embedding expression vectors corresponding to individual entity nodes embedded in a latent space according to an embodiment.

Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only, and may be changed and implemented in various forms. Therefore, the form actually implemented is not limited only to the specific embodiments disclosed, and the scope of the present specification includes changes, equivalents, or substitutes included in the technical idea described in the embodiments.

Although terms such as first or second may be used to describe various components, such terms should only be construed for the purpose of distinguishing one component from another. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element.

It should be understood that when an element is referred to as being “connected” to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle.

Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate that the described feature, number, step, operation, component, part, or combination thereof exists, but one or more other features or numbers, It should be understood that the presence or addition of steps, operations, components, parts, or combinations thereof is not precluded.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in this specification, it should not be interpreted in an ideal or excessively formal meaning. don't

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. In the description with reference to the accompanying drawings, the same reference numerals are given to the same components regardless of reference numerals, and overlapping descriptions thereof will be omitted.

Recently, studies are being conducted to explain the complex life activities of diseases and to predict the relationship between diseases and related substances. In various fields such as molecular biology and genomics, open source databases for each related substance have been established and are used for research to understand the complex phenomena of life activities.

Real world data (RWD) refers to information in the natural setting of routine practice, as opposed to experimental settings where variables are controlled. Because the workings of the phenomena of whole life are not fully understood and understood, tests that measure factors and phenomena may not be sufficiently clear to design controlled settings in the laboratory.

The use of real world data in biomedical research reporting can be leveraged with the growth of information systems. Recently, along with the popularization of electronic devices and the prevalence of systematic data collection, biomedical data can be accumulated in the form of an Electronic Medical Record (EMR). This allows a reasonable amount of data to be collected computationally. In clinical research, electronic medical record data demonstrates prediction of clinical events, medical prognosis, or prediction of biomedical relevance (e.g., drug repositioning). and can be used to support clinical regulatory surveillance decisions.

In the following, biomedical correlations using newly built biological networks based on the effective combination and combination of real world data and biomedical data are described. Biomedical correlation is an association between entities in a biological database, and may indicate whether biological characteristics are related to each other, and the association between entities exemplarily includes drug re-creation, drug side effect prediction, drug treatment effect prediction, disease occurrence prediction, and the like. can include Predicting previously unknown associations between drugs, diseases, prescriptions, and/or diagnoses can be referred to as drug repositioning. In this specification, drug re-creation is mainly described as an example of biomedical relevance prediction, but is not limited thereto.

1 shows a block diagram of a biomedical correlation prediction device according to an embodiment.

Computing device 100 according to an embodiment may include a processor 110 and a memory 120 .

The processor 110 may extract real world data including a diagnosis code and a prescription code linked to patient information from a real world database. The processor 110 may extract biomedical data including gene information, disease identifiers, chemical substance identifiers, and parent structure identifiers from a biomedical database. Processor 110 may build a new biological network by combining real world data and biomedical data. The processor 110 trains a machine learning model using the random walk data generated based on the random walk from the constructed new biological network, and based on the trained machine learning model, the embedding vector of the object node of the new biological network. Based on this, it is possible to predict links between entity nodes. Detailed operations of the processor 110 will be described with reference to FIGS. 2 to 12 below.

The memory 120 may store a machine learning model for biomedical relevance prediction. Memory 120 may temporarily or permanently store data required for biomedical relevance and training of machine learning models.

In addition, the computing device 100 may further include an output unit (eg, a display). The output unit may visually output an embedding vector mapped to the latent space and/or a link prediction result between entity nodes.

The computing device 100 according to an embodiment may map terms commonly used in the electronic medical record and the biomedical database, and build a database in the form of a graph. An effective graph structure capable of searching for potential benefits and harms of drugs is constructed in the graph database, and the computing device 100 can search for drug effects through graph embedding technology. The computing device 100 may search for drug indications by simultaneously utilizing the electronic medical record and the biomedical database.

The computing device 100 may perform mapping for terms commonly used in the electronic medical record and the biomedical database. The computing device 100 can effectively embed the latent representation of the graph when the electronic medical record and the biomedical database are combined and constructed in the form of a graph.

In the present specification, a biological network is any network applied to data about a biological system, and may represent a network capable of providing a mathematical expression about a connection between data. Biological networks can be represented as graph-structured data. A graph is a data structure composed of vertices and edges. In this specification, a vertex may also be referred to as an entity node. For example, biological data included in a biological network may include one or more entity nodes. An entity node is a node that is defined as various entity types according to entities included in the biological network database and indicates information corresponding to the entity type (eg, disease type indicates the type of disease). Exemplary entity types described include diagnosis type, prescription type, and patient type in the real world database, and gene type, disease type, chemical substance type, and parent structure type in the biomedical database. Entity nodes of each entity type are described in FIGS. 3 and 4 below. Each entity node may be connected to one or more other entity nodes through an edge. In this specification, an edge having no direction and weight is mainly described, but is not limited thereto.

For reference, in the present specification, a biological network includes a real world data network included in a real world database and a biomedical network included in a pre-built biomedical database. can do. Data extracted from the real world data network may be represented as real world data (RWD), and data extracted from the biomedical network may be represented as biomedical data.

2 is a flowchart illustrating a biomedical correlation prediction method according to an embodiment.

First, in step 210, the computing device may extract real world data including a diagnosis code and a prescription code associated with patient information from a real world database. For example, the computing device connects a patient node indicating patient information to a prescription node indicating a prescription code extracted corresponding to the patient information and a diagnosis node indicating a diagnosis code through an edge, thereby providing real world data in a graph structure. can create

In operation 220, the computing device may extract biomedical data including genetic information, disease identifier, chemical substance identifier, and parent structure identifier from the biomedical database. For example, the computing device may connect a gene node indicating genetic information to a disease node indicating a disease identifier corresponding to the genetic information through an edge. The computing device may connect the disease node to a chemical node indicating a chemical identifier corresponding to the disease identifier through an edge. The computing device may connect the chemical node to a parent structure node indicating a parent structure identifier corresponding to the chemical identifier through an edge. The computing device may generate biomedical data in a graph structure including a gene node, a disease node, a chemical node, and a parent structure node connected through an edge.

For reference, in biomedical data, a disease node may be connected to a parent structure node via a chemical node through an edge, and a chemical node may be connected to a gene node via a disease node through an edge.

Then, at step 230, the computing device may build a new biological network by combining real world data and biomedical data. For example, the computing device may build a new biological network by connecting a diagnosis node of real world data and a disease node of biomedical data, and a prescription node of real world data and a chemical node of biomedical data.

In operation 240, the computing device may train the machine learning model using random walk data generated based on the random walk from the constructed new biological network.

Subsequently, in step 250, the computing device may predict links between entity nodes based on the embedding vectors of entity nodes of the new biological network based on the trained machine learning model.

3 may represent real world data according to an embodiment.

The computing device may extract real world data 300 from a real world data network. The real world data network is a network composed of data accumulated in the form of an EMR (Electric Medical Record), and may also be referred to as a real world database. The real world data 300 may represent data in the real world that includes various variables, rather than clinical data in an experimental setting in which arbitrary variables are limited. Variables that affect life phenomena are very diverse, and experimental results obtained through limited experiments may differ from results obtained in the real world. Therefore, the real world data 300 can have a clinical basis because it contains various variables that are not limited. Accumulated real-world data 300, for example, electronic medical record data, may provide opportunities to better explain biological phenomena or actual clinical events through new approaches.

Electronic medical record data according to an embodiment may be extracted from the CardioNet database of Asan Medical Center (AMC). A structured entity of the real world data 300 may be extracted from a patient demographics table, a diagnosis table 310 (diagnosis table), and a medication table 320 (medication table). Patients in the real world database can be numbered anonymously by Patient Identifier (PAID). A patient encounter number (EN_NO) may be given at every first visit for different types of diagnoses. A key in which the patient identifier (PAID) and the patient encounter number (EN_NO) are string-coupled may be used to search the diagnosis table 310 and medication table 320 of the patient. The patient encounter number (EN_NO) may be replaced with a key. The patient encounter information may include a key in which a patient encounter number (EN_NO) and/or a patient identifier (PAID) and a patient encounter number (EN_NO) are string-coupled. For example, an entity node of a patient type (hereinafter referred to as 'patient node') is patient information, including a patient identifier node indicating a patient identifier (PAID) and a patient encounter indicating a patient encounter number (EN_NO) or key. Can contain nodes.

From the diagnosis table 310 (diagnosis table), the computing device obtains a diagnosis code (DICD), an in-date (INDT) of each key, and a first diagnosis date (The date) of each diagnosis code (DICD). of initial diagnosis, T_INDT) can be extracted. The diagnosis table 310 may include a diagnosis code (DICD) assigned based on the Korean Classification of Diseases (KCD). The diagnosis code (DICD) may be a code indicating a list of diseases and symptoms occurring at each visit to a patient. A diagnosis code (DICD) (e.g., a diagnosis code of an institution at Asan Medical Center) is a Korean disease created based on the "International Statistical Classification of Diseases and Related Health Problems (ICD)" It may be an extended code from the Korean Classification of Diseases (KCD). Diagnosis codes (DICD) are often originating from source code and institutional requests such as health insurance claims or service details (e.g., emergency visits). You can include code according to your needs. For example, an entity node of a diagnosis type (hereinafter referred to as 'diagnosis node') may be an entity node indicating a diagnosis code (DICD) indicating diagnosis contents.

The computing device may extract a prescription code (ODCD) from the medication table 320 according to an element column (ELEM) for each key. The medication table 320 may include an order code (ODCD) and ingredient names assigned based on RxNorm and RxNorm Extension. RxNorm and RxNorm Extension are US-specific terms for drugs that include all drugs available on the US market, and can also be used for personal health record applications. RxNorm is part of the Unified Medical Language System (UMLS) terminology. May be maintained at the US National Library of Medicine. The prescription code (ODCD) may be a code indicating information about medication administered to a patient visiting a hospital. For example, an object node of a prescription type (hereinafter referred to as a 'prescription node') may be an object node indicating a prescription code (ODCD) indicating contents of a prescription.

As described above, the prescription code (ODCD) and the diagnosis code (DICD) may include the patient's disease history and the type of medication administered according to changes in the patient's condition or characteristics of the patient upon hospitalization.

According to an embodiment, the computing device includes a prescription code (ODCD) and a diagnosis code (DICD) associated with patient information (eg, the patient identifier (PAID) and patient encounter number (EN_NO) described above) as objects in the real world. Data 300 can be extracted. The computing device may extract real world data 300 in which patient characteristics are reflected. For example, the computing device transmits a patient node indicating extracted patient information to a prescription node indicating an extracted prescription code (ODCD) corresponding to the corresponding patient node and a diagnosis node indicating a diagnosis code (DICD) through an edge. By connecting, real world data 300 in a graph structure can be created. In addition, the computing device may further connect an outcome node indicating the defined outcome 330 to the patient node via the edge.

In addition, the computing device may extract a column or table representing a defined outcome 330 for a specific use of the structure. For example, in FIG. 3 , a result indicating whether or not the patient's condition was positively improved and/or negatively deteriorated by the prescription of the drug corresponding to the prescription code (ODCD) may be included. can In another example, a computing device may determine the death date of a patient enrolled in cancer for each patient identifier (PAID) from the patient demographic table, as a defined outcome of patient adjustment for anticancer drugs. a cancer registration patient) (CDTH) can be extracted.

4 may represent biomedical data according to an embodiment.

According to an embodiment, the computing device may extract biomedical data 400 from a pre-built biomedical database. Biomedical databases may contain data that have manually curated relationships by experts. For example, the biomedical database may be a Comparative Toxicogenomic Database (CTD) database;

The CTD database may represent a comparative database where interactions of toxicogenomical relationships are manually curated. The CTD database provides data on the relationship between chemical mechanisms and genes of chemicals and between drugs and diseases. Relationships by entity type appearing in each dataset of the CTD database may be curated by experts. The CTD database may be an extended database to cover not only the chemical toxicity of environmental chemicals, but also the pharmacological activity of toxicogenomics. Furthermore, the curation of gene-related entities is broadened to phenotypes, gene ontology, and pathways, including genes, proteins, and metabolites. It can give the possibility of expanding to interdisciplinary research involving metabolites and their related reactions.

The disease identifier (Disease_ID) may be determined according to the “disease” branch of Medical Subject Headings (MeSH) of the National Library of Medicine (NLM) under the National Institutes of Health. In addition, the disease identifier (Disease_ID) may be determined according to Online Mendelian Inheritance in Man (OMIM) for some genetic disorders not covered by MeSH. For example, an object node of a disease type (hereinafter, a 'disease node') may be an object node indicating a disease identifier (Disease_ID) indicating a type of disease.

The gene information (Gene_Symbol) may include information about a toxicologically important gene symbol or protein symbol from a gene database of National Center for Biotechnology Information (NCBI). For example, an object node of a gene type (hereinafter referred to as 'gene node') may be an object node indicating gene information (Gene_Symbol) indicating the type of toxicologically important gene symbol and/or protein symbol. .

The chemical identifier (Chemical_ID) can be determined based on the “Chemicals and Drugs” branch of NLM MeSH. Also, the chemical identifier (Chemical_ID) may be determined according to molecular reagents, environmental chemicals, and clinical drugs. For example, an entity node of a chemical type (hereinafter referred to as 'chemical node') may be an entity node indicating a chemical identifier (Chemical_ID) indicating a type of chemical.

The computing device may generate and extract biomedical data 400 based on relationship data of the CTD database. Relational data is, for example, in the chemical-disease dataset 410, the disease-gene dataset 420, and the chemicals dataset 430. Based on the biomedical data 400 can be generated and extracted. Interactions (eg, relationships) between chemicals and diseases in the chemical-disease dataset 410, interactions between diseases and genes in the disease-gene dataset 420, and chemicals dataset 430 ) can be a manually curated dataset by experts based on published literature.

The chemical-disease dataset 410 is a dataset on interactions between chemicals and diseases, in which unnecessary environmental chemicals are removed, and components of drugs corresponding to the remaining chemicals, diseases, and chemical substances- The disease relationship may be extracted data. For reference, ingredients of a clinical drug may be sifted by mapping between a chemical substance identifier (Chemical_ID) and prescription information of an electronic medical record database.

The disease-gene dataset 420 corresponds to a disease in an electronic medical record database among disease-gene interactions to reflect the mechanism of a drug acting through a reaction with a target protein or target gene. to include curated disease-gene relationships. The computing device may extract genetic information of a predetermined number from among a plurality of genetic information related to a corresponding disease from a biomedical database. Based on the inference score provided by the biomedical database, the computing device may extract gene information related to the disease up to a predetermined number in descending order from gene information having the highest inference score for the corresponding disease. For example, the computing device may extract top 10 genetic information based on the inference score. In a graph-structured network without node weights, all associations (eg, relationships), regardless of differences in importance, should be evaluated with the same probability when performing a random walk, so in the biomedical data 400 The number of edges indicating associations between entities (eg, associations between entity nodes) needs to be normalized. Therefore, genetic information linked to the 10 most important associations among a plurality of genetic information related to a disease can be extracted to normalize the importance, and the probability of walking for important genes can be standardized across the network.

The inference score can be used to statistically evaluate functional relationships between proteins in Protein Protein Interaction (PPI). The inference score can be used to pick the top 10 genes associated with each disease or chemical in chemical-gene and disease-gene associations. The inference score can be derived from the concept of how biological networks innate to form scale free random networks. It can show the degree of similarity between chemical-gene-disease networks and similar scale-free random networks for assessing the degree of atypical connectivity. Furthermore, Common Neighbors methods such as the connectivity of a chemical/disease and a gene to the amount of related disease/chemical and the connectivity of the inferred chemical/disease itself can be used to calculate the inference score.

The chemicals dataset 430 may also be referred to as a chemical vocabulary dataset, and may be used to reflect reactions in which a drug differs from a parent structure to a functional group. The computing device may extract a chemical substance-parent structure relationship from the chemicals dataset 430, for example, a parent structure identifier (Parent_ID) linked to one chemical substance identifier (Chemical_ID). A chemical-parent structure relationship may include parental structure information that is linked to all descendants. Therefore, the computing device extracts the parent structure of a chemical substance and connects chemicals having different functional groups to reflect characteristics of drugs having similar chemical formulas but different effects due to differences in functional groups in the biomedical data 400. have. For example, an entity node of a parent structure type (hereinafter referred to as 'parent structure node') may be an entity node indicating a parent structure identifier (Parent_ID) indicating the type of parent structure.

Also, to combine toxicogenomical relationships, the disease-gene dataset 420 and the chemical-gene dataset can be used to do so.

In the example shown in FIG. 4 , the computing device includes biomedical data 400 including genetic information (Gene_Symbol), disease identifier (Disease_ID), chemical substance identifier (Chemical_ID), and parent structure identifier (Parent_ID) as entities from a biomedical database. ) can be extracted. The extracted biomedical data 400 includes gene information (Gene_Symbol), disease identifier (Disease_ID), disease identifier (Disease_ID), chemical substance identifier (Chemical_ID), chemical substance identifier (Chemical_ID), and parent structure identifier (Parent_ID) corresponding to each other. can include For example, the computing device may connect a gene node indicating the extracted gene information (Gene_Symbol) to a disease node indicating a disease identifier (Disease_ID) corresponding to the gene information (Gene_Symbol) through an edge. Similarly, the computing device connects the chemical node indicating the chemical identifier (Chemical_ID) corresponding to the corresponding disease identifier (Disease_ID) to the disease node through the edge, and connects the corresponding chemical identifier (Chemical_ID) to the chemical node. A parent structure node indicating a parent structure identifier (Parent_ID) corresponding to may be connected. Accordingly, the computing device may extract biomedical data 400 including gene information (Gene_Symbol), disease identifier (Disease_ID), chemical substance identifier (Chemical_ID), and parent structure identifier (Parent_ID) through edge connection between nodes. . In other words, the computing device may generate biomedical data having a graph structure including a gene node, a disease node, a chemical node, and a parent structure node connected through an edge. For example, the disease node may be connected to the parent structure node via the chemical node through an edge, and the chemical node may be connected to the gene node via the disease node through an edge. However, although the connection process has been described in an arbitrary order for convenience of explanation, the connection order is not limited to the above-mentioned one. Depending on the design, the aforementioned connection between nodes may be performed in the reverse order or in any order.

Since the patient's genomic data is lacking in the EMR, an edge may not be created. Relevant information may not be added to biomedical databases. Nonetheless, in the CTD database, diseases may be correlated to diseases, or chemicals may be correlated to chemicals, through self-statistical inference scores derived from statistical evaluation of common-neighbor methods.

5 to 7 illustrate mapping of real world data and biomedical data according to an exemplary embodiment.

According to an embodiment, the computing device connects the diagnosis node 512 of the real world data 510 and the disease node of the biomedical data 530, and the prescription node 511 of the real world data 510 and the biomedical data 530 A new biological network can be built by connecting the chemical nodes of For example, the computing device maps the diagnosis node 512 and the disease node corresponding to the same unique concept identifier based on a "Common Data Model" (CDM) for biomedical terminology, thereby real World data 510 and biomedical data 530 may be combined. The computing device may connect the prescription node 511 and the chemical node corresponding to the same unique concept identifier based on a common data model. In a real world database, for example, a clinical data set, in-hospital codes (eg, the above-mentioned diagnosis code (DICD) and prescription code (ODCD)) are generated according to the unique code system of each hospital, so that in-hospital To apply the code to the various sources, the computing device can perform mapping based on a common data model.

The common data model is a model for the thesaurus, ontology, and informatics of biomedical terminology, for example, Unified Medical Language System (UMLS 520) ), Mondo Disease Ontology (Mondo), and Observational Health Data Sciences and Informatics (OHDSI).

UMLS 520 may be a thesaurus of all biomedical domains designed to integrate vocabulary sources in an interoperable manner. The UMLS 520 assumes that a terminology string of a biomedical concept itself represents a meaning and has synonyms for the same meaning. In the UMLS 520, different vocabularies having the same meaning may be grouped into one Concept Unique Identifier (CUI). For example, within the same concept, synonyms and lexical variations each assigned to different lexical unique identifiers (LUIs) may be grouped into the same concept unique identifier. Even within the same synonyms, different string unique identifiers (SUIs) are given for string variants (eg, uppercase or lowercase letters), and the corresponding string variants can also be grouped into the same concept unique identifier. The UMLS 520 Metathesaurus may be a large-scale thesaurus of biomedical concepts (e.g., a thesaurus) where columns consist of names having biomedical concepts under defined categories.

Mondo is an ontology that merges multiple disease term domains and can represent an ontology that seeks to provide accurate mappings across lexical sources. UMLS 520, which provides concepts and related synonyms, may be capable of one-to-many mapping. On the other hand, Mondo can provide one-to-one mapping using "K-means Bayesian-ontology inference" (kBoom). In Mondo, a probabilistic ontology is created from source-provided mappings and can be merged into the ontology by prioritizing probabilities by kBoom. This high-quality one-to-one mapping can be utilized for networks for random walk-based graph embedding, and the number of links in a node can affect the quality of the embedding.

OHDSI can present standardized observational research that enables health data analysis between organizations at the global level. OMOP (Observational Medical Outcomes Partnership) CDM is a data model with the same structure and standard that can be applied to medical data of different structures possessed by each medical institution, and can be used to transform data from multi-institutional sources around the world. As countries have their own national health departments where healthcare products are approved, clinical drug products may differ from country to country. To standardize global drugs, OMOP CDM can provide an interoperable standardized vocabulary for medicine called RxNorm Extension. In other words, OHDSI can be used to classify drugs prescribed in individual countries into a global concept. RxNorm Extension may be a vocabulary having an extension of concepts derived from RxNorm, a US specific standard term. It may be possible to transfer local institution term drug data to global drug concept using OMOP CDM.

According to an embodiment, a concept unique identifier (CUI) corresponding to the prescription node 511 of the real world data 510 may be obtained based on string mapping and element mapping. A concept unique identifier (CUI) corresponding to the diagnostic node 512 of the real world data 510 may be obtained based on string mapping and source mapping.

An example using the UMLS 520 as a common data model is mainly described below, but is not limited thereto.

6 is a diagram illustrating a mapping operation between a disease node indicating a diagnosis code (DICD) using UMLS and a disease node indicating a disease identifier (Disease_ID).

Codes of real world data (eg, a diagnosis code (DICD) and a prescription code (ODCD)) according to an embodiment may be generated based on a concept name of a common data model. In the case where identifiers (IDs, identifiers) corresponding to different statement strings (eg, synonym unique identifier (LUI) and string unique identifier (SUI)) have the same concept in the common data model, Identifiers corresponding to different sentence strings may be grouped into the same concept unique identifier (CUI). Within the concept of the same concept unique identifier (CUI), sentence strings indicating various lexical synonyms may be included. A synonym unique identifier (LUI) may be assigned to each lexical variation, and a string unique identifier (SUI) may be assigned to each string variation of each lexical variation. For example, FIG. 6 shows a list of various sentence strings having “C0026919” as a concept unique identifier (CUI). Accordingly, the computing device may perform concept mapping using a string unique identifier of a string field of the common data model.

The computing device may extract a string corresponding to a code (eg, a diagnosis code (DICD) and a prescription code (ODCD)) indicated by an entity node (eg, a diagnosis node) of real world data. The computing device may extract a concept unique identifier (CUI) having a character string identical to a character string corresponding to a code of real world data from a common data model (eg, UMLS). Exemplarily, since a character string is the smallest unit of UMLS, the computing device may basically search for a concept unique identifier (CUI) in units of character strings, but is not limited thereto. The computing device may search for a unique concept identifier (CUI) having a code matching a code of real world data or a character string matching a character string corresponding to a code. The computing device may determine that the strings corresponding to the codes of the real world data and/or identifiers of the biomedical data match each other only when the strings of the common data model are exactly the same. In FIG. 6, mapping between a diagnosis code (DICD) and a disease identifier (Disease_ID) is mainly described.

The diagnostic code (DICD) may be based on the International Statistical Classification of Diseases and Related Health Problems (ICD) 10, which consists of a 'core' classification. ICD 10 can be a list of three character categories and a subdivision in the numeric character after the decimal point. Epidemic disease and severe disease were chosen as headings for the core categories and are mandatory for reporting, and up to 10 subdivisions may appear as numbers after the decimal point. Subcategories can be used for detail time and variety of groups of different sites or conditions. For example, 'A01.0' for typhoid fever and 'A20.0' for lymph node plague may be used. The 'A31.9' code shown in FIG. 6 may indicate an atypical mycobacterial infection (Not Otherwise Specified (NOS)). A diagnosis code (DICD) may be written as a 3-digit number as the disease details are often not recorded by physicians, with the numbers after the decimal point having extended meaning for institutions (e.g. hospitals). can be hierarchically represented.

According to an embodiment, the computing device may extract a first string corresponding to the diagnostic code (DICD) indicated by the diagnostic node. The computing device may search for a first concept unique identifier (CUI) having a first string extracted from the common data model. In the example shown in FIG. 6 , the computing device generates a concept unique identifier (CUI) from the column 631 of the UMLS having the same string 'Atypical mycobacterial infection NOS' as the diagnosis code (DICD) 'A31.9' of the diagnosis node 610. ) 'C0026919' can be searched. A computing device may obtain various representations of the diagnostic code (DICD) by extracting other codes and character strings having the same concept unique identifier (CUI). For example, the computing device includes other string unique identifiers (SUIs) belonging to the same concept unique identifier (CUI) (eg, 'S0016794', 'S0016795', etc. shown in FIG. 6) and other synonym unique identifiers (LUIs) ( For example, information 632 corresponding to 'L0026919', 'L10114762', and 'L6525836' shown in FIG. 6 may be obtained. The computing device may determine a disease identifier (Disease_ID) corresponding to a string unique identifier (SUI) belonging to the same concept unique identifier (CUI) and a synonym unique identifier (LUI) based on the UMLS MeSH code. The computing device may map a diagnosis node to a disease node indicating a disease identifier (Disease_ID) corresponding to the retrieved first concept unique identifier (CUI). Accordingly, the computing device may map disease nodes corresponding to synonyms and synonyms to one diagnosis node. In addition, when the brand name of the prescribed drug is used as the concept name, the synonym unique identifier (LUI) included in the concept unique identifier (CUI) corresponding to the brand name of the drug and the disease identifier (Disease_ID) corresponding to the string unique identifier (SUI) can be mapped.

About 61% of the diagnostic codes (DICD) are mapped by the MeSH terminology, and the remaining diagnostic codes (DICD) that are not mapped may be mapped to the disease identifier (Disease_ID) based on source mapping. For example, the computing device may map a disease identifier (Disease_ID) pre-mapped to a field defined in the disease ontology and Mondo for the remaining diagnostic code (DICD) after MeSH code mapping. For example, string mapping may be insufficient for Bell's palsy, but when source mapping is applied, the computing device can map to a code with the concept of facial paralysis as a symptom and familial inheritance.

In RxNorm of real world data, prescriptions are classified by dose, usage, and ingredients, etc., whereas in MeSH of biomedical data, drugs with the same ingredients may have the same MeSH concept code. Accordingly, since ingredient name conversion is required, the computing device may obtain ingredient name information corresponding to a prescription code (ODCD) of real world data and/or a chemical substance identifier (Chemical_ID) of biomedical data using the OHDSI data source.

The computing device according to an embodiment may perform mapping between a prescription code (ODCD) of real world data and a chemical substance identifier (Chemical_ID) of biomedical data similarly to the above description. A prescription code (ODCD) can be generated based on RxNorm/RxNormExtension. The RxNorm's vocabulary of prescription codes (ODCD) can contain complex concepts because they are classified by dose, usage, ingredient, etc. On the other hand, in MeSH, drugs with the same ingredients are indicated by the same MeSH concept code, so ingredient name conversion may be required. For reference, as shown in FIG. 7 , the computing device may extract the component name of RxNorm/RxNormExtension through element mapping using the 'has_ingredient' relationship 700 of OHDSI.

The computing device according to an embodiment may extract a second string corresponding to the prescription code (ODCD) indicated by the prescription node. For example, the computing device may extract the ingredient name and drug name of each drug according to the prescription code (ODCD). The computing device preprocesses (eg, simple Natural Language Processing (NLP)), case conversion, position change of prepositions, spacing correction for the ingredient name and/or drug name of the drug obtained in correspondence with the prescription code (ODCD). , Parentheses processing including explanation) may be performed to extract the second string converted to lowercase letters.

The computing device may search for a second concept unique identifier (CUI) corresponding to the second string extracted from the common data model. For example, the computing device may assign an RxNorm code based on the RxNormExtension of OHDSI for the ingredient name, and then map it to the UMLS MeSH code, which is a common data model. In other words, the computing device may map the same MeSH code as RxNorm having the same components to RxNormExtension having the same components but different usage and capacity from RxNorm. The computing device may obtain a second concept unique identifier (CUI) of the UMLS for each ingredient name (eg, a second string) by string-mapping the ingredient name of the prescription code (ODCD) to the UMLS.

The computing device may map the prescription node to a chemical substance node indicating a chemical substance identifier corresponding to the retrieved second concept unique identifier (CUI). The computing device may extract entity nodes having the same second concept unique identifier (CUI) for the entire vocabulary in the UMLS to find another representation of the ingredient name corresponding to the prescription code (ODCD). Similar to the above, the computing device maps chemical nodes corresponding to synonym unique identifiers (LUIs) and string unique identifiers (SUI) belonging to the same second concept unique identifier (CUI) to prescription nodes, thereby providing biomedical data and real world data. data can be combined.

8 illustrates a mapping result between real world data and biomedical data according to an embodiment.

A computing device according to an embodiment shows new biomedical data including real world data 810 and biomedical data 820 combined by the mapping described above with reference to FIGS. 5 to 7 . A new set of biomedical data may constitute a new biological network as described below with reference to FIGS. 9 and 10 . Real world data 810, biomedical data 820, new biomedical data, and new biological networks can be implemented in a graph structure.

Real world data 810 may include outcome node 815, patient identifier node 811, patient encounter node 812, prescription node 813, and diagnosis node 814, as described above. The real world data 810 includes an edge connecting the outcome node 815 and the patient identifier node 811, an edge connecting the patient identifier node 811 and the patient encounter node 812, and a patient encounter node 812 and the patient encounter node 812. An edge connecting the prescription node 813 and an edge connecting the patient encounter node 812 and the diagnosis node 814 may be included.

As described above, the biomedical data 820 may include a chemical node 821, a parent structure node 822, a disease node 823, and a gene node 824. The biomedical data 820 includes an edge connecting the chemical node 821 and the parent structure node 822, an edge connecting the chemical node 821 and the disease node 823, and the disease node 823 and the gene node. 824 may be included.

The new biomedical data may include both object nodes and edges of the real world data 810 and the biomedical data 820 described above. In addition, the new biomedical data may further include edges newly generated by the mapping described above with reference to FIGS. 5 to 7 . For example, the new biomedical data may include an edge connecting the prescription node 813 and the chemical node 821 and an edge connecting the diagnosis node 815 and the disease node 823 . The edge connecting the prescription node 813 and the chemical node 821 and the edge connecting the diagnosis node 815 and the disease node 823 are cyclic when random walk data described later in FIG. 11 is generated. One part can be avoided. Cycle refers to the case where the starting node and passing node of the random walk are the same, and if cycles frequently occur in the random walk, performance degradation of the vector embedding model described later may occur, so it is necessary to prevent cyclics.

However, new biological data is not limited to what is shown in FIG. 8 , and patient nodes (eg, patient identifier node 811 and patient encounter node 812) may be excluded after vector embedding described later in FIG. 11 . have.

9 and 10 illustrate a new merged network created according to one embodiment.

In FIG. 8, new biomedical data generated by combining heterogeneous data is described, and in FIG. 9, connections between biomedical data are described. As shown in FIG. 9 , first real world data 911 may be connected to first biomedical data 921 . In addition, the first biomedical data may be connected to other second real world data 912 through a disease identifier (Disease_ID) and a chemical substance identifier (Chemical_ID) and third biomedical data 923 through a parent structure. The third biomedical data 923 may be connected to the third real world data 913 through a disease identifier (Disease_ID) and a chemical substance identifier (Chemical_ID). As shown in FIG. 9 , the computing device may construct a new biological network by connecting a plurality of entity nodes to one entity node through an edge.

10 illustrates an example of a newly constructed biological network 1000 . Exemplarily, the number of object nodes and edges of the newly constructed biological network 1000 may be shown in Table 1 below.

Datasets Entities and relationships EMR
Entities PAID KEY DICD ODCD 360409 14967169 1025 4798 Relationships
(edge) PAID-KEY KEY-ODCD KEY-DICD KEY-CDTH 14967169 41730466 22219689 360409 CTD
Entities DiseaseID ChemicalID GeneSymbol ParentID 7643 1805 1496 733 Relationships
(edge) DiseaseID-ChemicalID DiseaseID - GeneSymbol ChemicalID - ParentID ChemicalID - GeneSymbol 28580 4361 1805 282409 Mapping Relationships
(edge) ODCD-Chemical ID DICD-DiseaseID 6633 7989

11 illustrates training of a machine learning model and biomedical correlation using the trained machine learning model according to an embodiment. According to an embodiment, a computing device generates a random walk based on a random walk from a new biological network constructed. A machine learning model may be trained using the work data 1120 .

The computing device may generate random walk data 1120 based on the random walk from the new biological network constructed in FIG. 10 . In this specification, a graph of a new biological network integrated with real world data (eg, electronic medical record data) can be defined as G = (ν, ε). Here, the node set ν is new biological data, and includes object nodes 1110 of 5 entity types of real world data and object nodes 1110 of 4 entity types of biomedical data, a total of 9 heterogeneous entity nodes 1110 ( heterogeneous entity nodes). Exemplarily, the set of entity nodes of the first entity type ν ₁ = {ν _{11 , ...,} ν _1n }, to the set of entity nodes of the ninth entity type ν ₉ = {ν _{91 , ...,} ν _9n } can be expressed as The node set v may include a total of V node entities, where n and V may be integers greater than or equal to 1. Entity node sets may be generated from an adjacency matrix representing adjacencies between entities. For example, each row and each column of the adjacency matrix represents the i-th entity and the j-th entity, and the value of the element in the j-th column of the i-th row of the adjacency matrix is the adjacency between the i-th entity and the j-th entity (e.g., adjacency). Illustratively, v _ij may indicate the object node 1110 of the jth instance of the ith entity type. Edge ε is an edge connecting entity nodes 1110 ν _ij and may represent association and/or relationship in a data set without weight or direction. An edge can be expressed as ε ⊆V×V.

The random walk data 1120 may be data representing a random sequence based on a random walk between neighboring entity nodes. The random walk for the graph-structured data network is an entity node (1110) randomly selected from among one or more entity nodes (1110) connected from an arbitrary entity node (1110) constituting the data network to the corresponding entity node (1110) through an edge. ) may indicate an operation of selecting and extracting. The length of the random walk data 1120 may be determined according to the length of the random walk, that is, the number of trials of the random walk. The direction and probability of the random walk may be determined according to the direction and weight of the edge. As described above, in this specification, for convenience of description, it is assumed that there is no direction and all weights are the same. The random walk data 1120 may also be referred to as a random walk input.

The random walk input is data generated to capture topological information of the graph network, and may also be referred to as 2 ^nd order random walk data. Similarity between entity nodes included in the new biological network may be defined as an occurrence probability of a neighboring entity node 1110 in a given feature representation, eg, an embedding vector. The probability of occurrence of the neighboring object node 1110 may also be interpreted as the probability that an edge exists between the neighboring object nodes 1110. The computing device may generate a random walk set N _S (u) for graph embedding. Starting from an initial node u through a random walk, the computing device may generate a walk to a neighboring node through a predetermined random walk length using a random walk strategy S. Here, u may be u ∈ ν. As will be described later, within the random walk set N _S (u) represented by the random walk input, the probability of occurrence of neighboring nodes can be estimated based on the temporarily output embedding vector, and optimized for the similarity of the embedding space, A machine learning model 1140 described below may be trained.

For example, the computing device may select an initial object node u in the newly constructed biological network. The computing device may select an entity node according to an individual random walk by sequentially performing a random walk as much as a predetermined random walk length from the selected initial entity node u. The computing device may generate random walk data 1120 indicating a random walk set N _s (u) including an initial object node u and an object node selected based on the random walk.

The random walk length may have a greater value than the total number of entity types of entity nodes constituting the new biological network. For example, in FIG. 11, an example in which the random walk length is 10 is described. For example, in generating a random walk, the walk length can be set to 10, and 380 different combinations of entity types can be created. For example, as the random work set N _s (u), [ν _1n , ν _2n , ν _1n , ν _2n , ν _1n , ν _2n , ν _1n , ν _2n , ν _1n , ν _2n ] to [ν _9n , ν _8n , ν _9n , ν _8n , ν _9n , ν _8n , ν _5n , ν _7n , ν _9n , ν _8n ] may be generated. For another example, in generating a random walk, approximately 40 combinations of 4 entity types and each entity node 1110 may be generated.

For reference, the return parameter p and the in-out parameter q may be determined to guide a search for sampling strategies such as breadth-first sampling (BFS) and depth-first sampling (DFS). To add a search bias to the transition probability of an edge, the previous node t is remembered, and the search bias can be calculated depending on its relative position compared to node t. For the next node x, the values of the shortest path between node t and node x are fixed as d _tx = {0,1,2}, so that the search bias α _pq (t, x) on edge (v,x) is calculated It can be. The search bias α _pq (t, x) may be determined as in Equation 1 below.

The reciprocal of p can be applied when moving closer to the previous node, and the reciprocal of q when moving away from the previous node. In Equation 1 above, since the new biological network exemplarily has edges with return parameter p = 1 and in-out parameter q = 1, the transition probability is independent of d _tx , α _pq (t, x) = 1 can be set to

The computing device may train the machine learning model 1140 using the aforementioned random walk input. The machine learning model 1140 is a model designed and trained to project each entity node 1110 indicated by a random walk input into an embedding vector in a latent space, and is, for example, a neural network model. can include In this specification, a "shallow neural network" including one hidden layer is mainly described, but is not limited thereto. The hidden layer of the machine learning model 1140 is responsible for an operation called a lookup table, and may also be referred to as a projection layer. The machine learning model 1140 may be designed as a skip-gram structure using negative sampling. However, the structure of the machine learning model 1140 is not limited, and may be implemented in various structures such as a convolutional neural network (CNN) and a recurrent neural network (RNN). In addition, the machine learning model 1140 includes a graph embedding model embedding data of a graph structure, and may include, for example, deep walk, node2vec, gcn, and the like.

As shown in FIG. 11 , the computing device may input input data 1130 corresponding to each entity node 1110 of the random walk input into the machine learning model 1140 . The computing device may input a one-hot vector corresponding to each of a plurality of entity nodes included in the random walk input to the machine learning model 1140 as input data 1130 . In FIG. 11, as an example, the random work input of [ν _9n , ν _8n , ν _9n , ν _8n , ν _9n , ν _8n , ν _5n , ν _7n , ν _9n , ν _8n ,] as the random work set _NS (u) is taken as an example. listen and explain Illustratively, the dimension of the input data 1130 (eg, one-hot vector) input to the machine learning model 1140 may be the same as the number of entity nodes of the new biological network 1000 generated in FIG. 10 described above. . When the total number of entity nodes constituting the new biological network 1000 is V, the dimension of the one-hot vector of the input data 1130 may also be V. Here, V may be an integer of 2 or greater. Based on the random walk input, the computing device may generate a one-hot vector indicating each entity node 1110 included in the random walk input. A one-hot vector indicating one object node 1110 may represent a vector in which an element value indicating a corresponding object node 1110 is '1' and an element value indicating other nodes is '0'. In FIG. 11, the computing device obtains a one-hot value in which the element value corresponding to the entity node ν _9n in the input data 1130 is '1' and the other element values are '0' for the random walk set N _S (u) among the random walk inputs. A one-hot vector in which only the element value corresponding to the vector, entity node ν _8n is '1', and a one-hot vector in which only the element value corresponding to the entity node ν _5n is '1' can be created. The computing device may train the machine learning model 1140 using the generated one-hot vector.

When the machine learning model 1140 includes the projection layer f, the parameter, eg, a weight matrix W between the projection layer f and the input layer, may be a VХd-dimensional matrix. Here, d may be an integer of 1 or greater. In other words, the computing device may generate a d-dimensional embedding vector by projecting a V-dimensional one-hot vector onto a d-dimensional one through the projection layer f of the machine learning model 1140 . For example, the computing device uses the machine learning model 1140 from one-hot vectors corresponding to the random work set N _S (u) to output data f(v _9n ), f(v _8n ), and f( v _5n ). As described above, the output data f(v _9n ), f(v _8n ), and f(v _5n ) are M-dimensional embedding vectors, and may represent the vector coordinates of each entity node 1110 in the d-dimensional latent space. have.

As described above, the machine learning model 1140 is a skip-gram model, and outputs an embedded representation for calculating a relationship between entity nodes for each entity node 1110, Sequences based on a random walk can be learned. Learning and/or training of the machine learning model 1140 may be performed based on an objective function value 1109 described later. Graph structure data G corresponding to the biological network 1000 described above in FIG. 10 may be expressed as G = (ν, ε). The projection layer f is a mapping function and can be expressed as f: v → R ^d . Here, R may represent parameters of a V×d-dimensional weight matrix that maps a node v to a d-dimensional vector space (eg, a latent space and/or an embedding space). The probability of observing a node, that is, the probability of existence of an edge connecting two nodes can be calculated by maximizing the problem shown in Equation 2 below.

Equation 2 described above may be interpreted as an objective function value 1109 that maximizes a connection probability between individual nodes included in the random walk set N _s (u). To simplify Equation 2 above, it can be assumed that the probability of occurrence of each node is independent of the probability of occurrence of other nodes, and that the two pairs of nodes have symmetrically equal effects on each other. Under the above-mentioned assumption and approximation using negative sampling, the objective function value 1109 of Equation 2 can be simplified as Equation 3 below as a dot product of features.

Equation 3 described above may be used to optimize the parameter f by a stochastic gradient ascent that maximizes Equation 2 representing the similarity of nodes and/or the probability of edge existence between nodes in the embedding space.

As described above, the computing device may calculate temporary embedding data by applying the machine learning model 1140 to the input data 1130 based on the random walk data 1120 . The temporary embedding data is the machine learning model 1140 before training is completed, that is, the embedding extracted from the input data 1130 (eg, a one-hot vector pointing to an arbitrary entity node 1110) using the temporary model. vectors can be represented. The computing device may update parameters of the machine learning model 1140 so that a connection probability between individual nodes in the random walk set N _s (u) is maximized based on the calculated temporary embedding data. Therefore, the computing device trains the machine learning model 1140 on the edge relationship obtained between the entity nodes in the new biological network 1000 through a random walk, and thereby vector coordinates in the latent space of the entity nodes connected to each other through edges. Parameters of the machine learning model 1140 may be updated so that are output adjacent to each other. Here, the projection layer f of the machine learning model 1140 may be a V×d-dimensional weight matrix, and a row vector corresponding to each row of the weight matrix represents a corresponding entity node 1110. It may be an embedding vector expressed. In other words, the kth row vector among the V plurality of row vectors of the weight matrix may be an embedding vector expressing the kth entity node 1110 among the V entity nodes of the new biological network 1000 . Here, k may be an integer greater than or equal to 1 and less than or equal to V.

The computing device predicts a link using the above-described machine learning model 1140 for a new biological network 1000 from which a certain percentage (eg, 10%) of edges are randomly removed, thereby generating a machine learning model 1140. ) performance can be verified.

12 illustratively illustrates embedding expression vectors corresponding to individual entity nodes embedded in a latent space according to an embodiment.

Exploring new indications for drugs can be referred to as drug re-creation or drug re-creation. A new indication of a drug may mean a new indication (eg, disease) different from the existing indication of a drug on the market, or may mean selection of a new target patient group for a drug that has failed in a clinical trial. New drug re-creation has the advantage of lowering the failure rate at a lower cost and being able to launch more quickly than the existing new drug development process.

In the new biological network 1000 constructed in FIG. 10 described above, entity nodes may be connected through edges according to relevance. In particular, a chemical node connected to a disease node may be interpreted as having an indication for a disease corresponding to a disease node with a chemical corresponding to the corresponding chemical node. According to one embodiment, the new biological network 1000 includes edges representing direct or indirect connections with other entity nodes such as parent structure nodes, prescription nodes, and diagnosis nodes, in addition to direct edges between chemical substance nodes and disease nodes. , the machine learning model 1140 described above in FIG. 11 may be a model designed and trained to predict an undiscovered missing edge in the new biological network 1000 .

According to an embodiment, an embedding vector corresponding to each of all entity nodes of the new biological network 1000 may be mapped to vector coordinates in the latent space 1200 . It may be determined that an edge exists between object nodes adjacent to each other in the latent space 1200 . For example, in FIG. 12 , when the distance between the first object node 1210 and the second object node 1220 in the latent space 1200 is less than a critical distance, the first object node 1210 and the second object node 1210 It may be determined that an edge 1290 exists between nodes 1220 . Therefore, the computing device may search potential edges in the new biological network 1000 according to vector distances between embedding vectors corresponding to object nodes in the latent space 1200, in addition to the explicit edges identified by the pre-built database. can

For example, the computing device may extract a first embedding vector corresponding to a first node (eg, a target disease node) among entity nodes of the new biological network from the machine learning model. The computing device may extract a second embedding vector corresponding to a second node (eg, a target chemical node) among entity nodes of the new biological network from the machine learning model. The computing device may determine whether an edge exists between the first node and the second node based on the extracted first and second embedding vectors. When the difference between the embedding vectors is small, it can be predicted that a potential edge and/or link exists between the two embedding vectors when the embedding vectors are similar. The computing device may determine that an edge exists between the first node and the second node in response to a case where a vector distance difference between the extracted first and second embedding vectors is less than a threshold distance value. Here, the vector distance between the two embedding vectors may be a Euclidean distance equal to the L2 norm. As described above, the computing device may predict a biomedical correlation by predicting a potential link based on a distance between embedding vectors in a latent space even if a link does not exist in a pre-constructed biological network.

For reference, in the above example, if the first node and the second node are each disease-drug, the computing device performs drug re-creation. When the first node and the second node become drug-drug, the computing device performs drug-drug interaction prediction. When each of the first node and the second node becomes a disease-disease, the computing device performs a disease-disease correlation search. When the first node and the second node respectively become a patient-disease, the computing device predicts the onset of a new disease of the target patient.

For example, the computing device may search for a chemical node less than a threshold distance based on a first embedding vector corresponding to a disease node indicating an arbitrary target disease. Thus, a drug having an effect on a target disease can be searched for. For another example, the computing device may search for a disease node less than a threshold distance value based on a second embedding vector corresponding to a chemical node indicating an arbitrary target chemical substance. Therefore, a disease corresponding to a new indication for the target chemical can be searched for.

The embodiments described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, the devices, methods and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate (FPGA). array), programmable logic units (PLUs), microprocessors, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. The device can be commanded. Software and/or data may be any tangible machine, component, physical device, virtual equipment, computer storage medium or device, intended to be interpreted by or provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on computer readable media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination, and the program instructions recorded on the medium may be specially designed and configured for the embodiment or may be known and usable to those skilled in the art of computer software. may be Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler.

The hardware device described above may be configured to operate as one or a plurality of software modules to perform the operations of the embodiments, and vice versa.

As described above, although the embodiments have been described with limited drawings, those skilled in the art can apply various technical modifications and variations based on this. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

Claims (14)

In the biomedical association prediction method performed by a computing device,
extracting real world data including a diagnosis code and a prescription code linked to patient information from a real world database;
extracting biomedical data including genetic information, disease identifiers, chemical substance identifiers, and parent structure identifiers from a biomedical database;
constructing a new biological network by combining the real world data and the biomedical data;
training a machine learning model from the constructed new biological network; and
Predicting links between entity nodes based on an embedding vector of entity nodes of the new biological network based on the trained machine learning model.
Biomedical relevance prediction method comprising a.
According to claim 1,
The step of extracting the real world data,
Generating the real-world data in a graph structure by connecting a patient node indicating patient information to a prescription node indicating a prescription code extracted corresponding to the patient information and a diagnosis node indicating a diagnosis code through an edge.
Biomedical relevance prediction method comprising a.
According to claim 1,
Extracting the biomedical data,
connecting a gene node indicating genetic information to a disease node indicating a disease identifier corresponding to the genetic information through an edge;
connecting the disease node to a chemical node indicating a chemical identifier corresponding to the disease identifier through an edge;
connecting the chemical node to a parent structure node indicating a parent structure identifier corresponding to the chemical identity through an edge; and
Generating biomedical data of a graph structure including the gene node, the disease node, the chemical node, and the parent structure node connected through an edge.
Biomedical relevance prediction method comprising a.
According to claim 3,
In the biomedical data, the disease node is connected to the parent structure node via the chemical node through an edge, and the chemical node is connected to the gene node via the disease node through an edge,
A method for predicting biomedical relevance.
According to claim 1,
The step of constructing the new biological network,
constructing the new biological network by connecting a diagnosis node of the real world data and a disease node of the biomedical data, and connecting a prescription node of the real world data and a chemical node of the biomedical data.
Biomedical relevance prediction method comprising a.
According to claim 5,
The step of constructing the new biological network,
mapping the diagnosis node and the disease node corresponding to the same unique concept identifier based on a common data model of biomedical terminology; and
Connecting the prescription node and the chemical node corresponding to the same unique concept identifier based on the common data model
Biomedical relevance prediction method comprising a.
According to claim 5,
The step of constructing the new biological network,
extracting a first character string corresponding to a diagnosis code indicated by the diagnosis node;
Retrieving a first concept unique identifier having the extracted first string from a common data model;
mapping the diagnosis node to the disease node indicating a disease identifier corresponding to the retrieved first concept unique identifier;
Biomedical relevance prediction method comprising a.
According to claim 5,
The step of constructing the new biological network,
extracting a second string corresponding to a prescription code indicated by the prescription node;
Retrieving a second concept unique identifier corresponding to the extracted second string from a common data model; and
Mapping the prescription node to the chemical substance node indicating a chemical substance identifier corresponding to the retrieved second concept unique identifier.
Biomedical relevance prediction method comprising a.
According to claim 1,
Training the machine learning model,
selecting an initial object node from the constructed new biological network;
selecting entity nodes according to individual random walks by sequentially performing random walks as long as a predetermined random walk length from the selected initial entity nodes; and
Generating random walk data indicating a random walk set including the initial entity node and an entity node selected based on the random walk
Biomedical relevance prediction method comprising a.
According to claim 9,
Training the machine learning model,
calculating temporary embedding data by applying the machine learning model to input data based on the random walk data; and
Updating parameters of a machine learning model to maximize a connection probability between individual nodes in the random walk set based on the calculated temporary embedding data
A method for predicting biomedical relevance further comprising a.
According to claim 1,
Predicting the link between the entity nodes,
extracting a first embedding vector corresponding to a first node among entity nodes of the new biological network from the machine learning model;
extracting a second embedding vector corresponding to a second node among entity nodes of the new biological network from the machine learning model; and
Determining whether an edge exists between the first node and the second node based on the extracted first embedding vector and the extracted second embedding vector
Biomedical relevance prediction method comprising a.
According to claim 11,
The step of determining whether the edge exists,
Determining that an edge exists between the first node and the second node in response to a case where a vector distance difference between the extracted first embedding vector and the second embedding vector is less than a threshold distance value
Biomedical relevance prediction method comprising a.
A computer program stored in a computer readable recording medium in order to execute the method of any one of claims 1 to 12 in combination with hardware.
In a computing device,
a memory storing machine learning models for predicting biomedical relevance;
Extracting real-world data including diagnosis codes and prescription codes linked to patient information from real-world databases, extracting biomedical data including genetic information, disease identifiers, chemical substance identifiers, and parent structure identifiers from biomedical databases, Constructing a new biological network by combining real world data and the biomedical data, training a machine learning model from the constructed new biological network, and embedding vectors of entity nodes of the new biological network based on the trained machine learning model. A processor that predicts links between entity nodes based on
Computing device comprising a.

Images